Beyond the Prompt: Why Creative Directors Are Training Their Own Custom Models

This is an illustrative feature. The agency, its people and the project below are fictional composites built to explain a real trend in creative work; no real studio, client or tool is described, and no figures are quoted.

The problem with the generic look

Imagine Atelier Ostrava, a fictional six-person creative studio. For a year the team used general-purpose image generators for moodboards and early concepts. The results were quick and often lovely, and increasingly alike. "Everything started to look like everything," says Mirela, the studio's imaginary creative director. "Clients noticed before we did."

That feeling, a sameness that comes from everyone drawing on the same well, is what pushes some teams beyond prompting. Prompting is a way of steering a model that belongs to someone else. Training your own, even a small one, is a way of building a tool that carries your taste.

What "custom model" means in practice

For most small studios, a custom model is not trained from scratch. It is a modest adjustment of an existing open model, using a few hundred carefully chosen images, so that it leans toward a particular palette, line quality or composition. The model stays small enough to run on a studio workstation.

In our fictional scenario, Atelier Ostrava assembles a dataset entirely from its own archive: illustrations the team drew, photographs they shot and textures they scanned. They document who made each piece, under what agreement, and whether it may be used for training. Anything unclear stays out.

The unglamorous centre: consent and provenance

The technical steps fit on a whiteboard. The hard questions do not.

• Who owns the training material? If freelancers contributed, did their contracts mention machine learning?
• Do clients allow it? Work made for a client is not automatically fair game.
• Can the model reproduce individual works? Small datasets can lead to near-copies, which is a legal and ethical risk.
• What happens to a person's style? Training on a living illustrator's work without agreement is a line many studios refuse to cross.

A sensible studio writes the rules down before a single image is fed in, and revisits them when something changes. That document is as much a part of the project as the model file.

The model is the easy bit. The hard bit is being able to say, out loud, where every training image came from. — Mirela, a fictional creative director in this illustrative scenario

How the workflow changes

With a house-style model in place, the studio's process shifts in three ways. First, early exploration gets faster, because outputs already sit near the brand's territory. Second, art directors spend more time curating and less time wrestling with prompts. Third, human illustrators move toward the parts of the job a model cannot do: deciding what a campaign should feel like, inventing the odd, specific detail that makes an image memorable, and finishing by hand.

Not everyone in the imaginary studio is thrilled. Dario, a junior illustrator, worries that the model will replace the loose sketching he is paid to do. The studio's answer, and the one we think is the right instinct, is to make the policy explicit: the model supports concept work, final art stays human-made, and any use in client deliverables is disclosed.

Treat the model like an asset

A house-style model should be handled like any other creative asset.

1. Version it. Keep a changelog of the dataset and settings, so a result can be traced.
2. Review it. Test regularly for unwanted artefacts, bias and near-duplicates of source works.
3. Retire it. When a licence expires or a contributor withdraws, remove their material and retrain.
4. Label its output. Tell clients and audiences when synthetic imagery is used.

Is it worth it?

For a studio that produces a steady volume of concept work in a distinctive style, perhaps. For one that takes on varied commissions, a custom model may be a costly distraction. The honest answer is to run a small pilot: one project, one narrow style, one agreed review date. Measure not only speed but also whether the results make the work better, and whether the team feels more or less ownership of what they make.

It is also worth asking what the alternative is. Some studios will decide that the lowest-risk, highest-integrity route is to skip generated imagery entirely, and lean on craft as their differentiator. That is a legitimate choice, and increasingly a marketable one.

The takeaway

Going beyond the prompt is less about technology than about authorship. A custom model can help a team keep its voice in a world of generic output, but only if the material it learns from is clearly theirs, consented and documented. Start small, write the rules first, keep humans in charge of the final image and be open with clients about what was made how.